Forecasting

Local energy markets require various types of forecasting. Even if the existing methods are more and more accurate, there is a continuous search for more advanced methods able to quantify the uncertainty of various electrical and price signals. Although a wide range of machine learning methods has been applied to electricity forecasting, in this chapter we will pass from linear models to state-of-the-art deep learning methods in an attempt to understand which are their most interesting challenges and limitations. The day-ahead electricity load forecast performance is analyzed for five EU countries. Consequently, we perform a comparison between Ordinary Least Squares, Ridge Regression, Bayesian Ridge Regression, Kernel Ridge Regression, Support Vector Regression, Nearest Neighbors Regression, Gaussian Process, Decision Trees, AdaBoost, Random Trees, and dense Multilayer Perceptron (MLP). Moreover, in an attempt to have accurate and fast methods with good generalization power, we introduce sparse neural networks and sparse training methods for electricity forecasting through the means of sparse MLPs trained with the Sparse Evolutionary Training procedure (SET-MLP).